Key takeaways
- ETL transforms data on the way in, while ELT lands raw data first and transforms it inside the warehouse.
- Cloud warehouses made ELT the default because storage is cheap and SQL transforms run where the data already sits.
- ETL still fits when sensitive fields must be removed before data leaves its source, or when the target cannot run transforms.
- Keep business logic out of the ingest tool, so switching pipelines later does not change a single number.
- Define metrics once in a warehouse semantic layer that BI tools, analysts and AI agents all read.
- 01Sources
CRM, billing, support, product events
- 02Extract and load
Managed ingest lands raw tables on a schedule
- 03Transform in the warehouse
SQL models, often in dbt
- 04Metric layer
Certified definitions for reports and agents
What is the difference between ELT and ETL?
Both patterns extract data from source systems, transform it into a usable shape and load it into a destination. The difference is the order. ETL transforms data in a separate engine before loading the finished tables. ELT loads raw data as it arrives and runs the transforms afterwards, inside the warehouse, usually in SQL.
ETL grew up when warehouse storage and compute were expensive, so it made sense to land only clean, aggregated tables. ELT became common with cloud warehouses such as Snowflake, BigQuery and Databricks, which separate storage from compute and can run large SQL transforms on demand.
Why do cloud warehouses usually favour ELT?
Landing raw data first keeps the full history of what each source sent. When a definition changes, you rebuild the model from the raw tables instead of asking the pipeline to re-extract months of data. Analysts can also read the raw tables when a number looks wrong, which makes debugging much faster.
Transforms written in SQL inside the warehouse are easier to review than logic hidden in a pipeline tool. Tools such as dbt keep those transforms in version control, run tests on them and record lineage between models, so a change goes through the same review as application code.
What does each pattern cost, and who maintains it?
Each pattern puts the cost in a different place. With ELT you pay to store raw data and to run transforms in the warehouse, and you pay your ingest tool on whatever unit it charges. Fivetran, for example, charges on monthly active rows, which it defines as the distinct rows synced to your destination in a calendar month. With ETL you pay for the transform engine and the people who maintain it.
Ownership matters as much as the invoice. In ELT, the data team owns the transforms in SQL, and anyone with warehouse access can read them. In ETL, the logic often sits in a tool only one or two engineers know, which becomes a risk when they move on.
When is ETL still the right choice?
ETL is still the better fit in a few situations, and there is no prize for forcing ELT onto them.
- Regulated fields, such as health or payment data, must be removed or masked before they leave the source environment.
- The destination is an application or operational database that cannot run transforms itself.
- Raw volumes are so large that landing everything would cost far more than the few fields you need.
- An existing ETL estate works, is documented and has owners, and the migration would bring no new capability.
Where should business metrics be defined?
Many ingest tools now offer transformations, and some ship prebuilt models for common sources. They are convenient for getting a first report running. The trouble starts when net revenue, active customer or churn are defined inside the pipeline. Switching ingest tools then means rebuilding those definitions, and the number on the board pack can change because a vendor contract ended.
Put metric definitions in the warehouse, in a semantic layer that BI tools, analysts and AI agents all read. The ingest tool moves data, the warehouse holds the models, and the semantic layer holds the agreed meaning of each number. Our post on semantic layers covers how to set one up.
What do AI agents need from the pipeline?
Agents need predictable tables with descriptions, certified metrics and access rules they cannot bypass. ELT helps, because the raw tables and the models sit in one place the agent can query through a governed role. An agent that has to guess which of three revenue columns is correct will give different answers to the same question, even with a clean pipeline.
How does Rig do it?
Rig Ingest runs scheduled syncs from more than 30 business tools into Snowflake, BigQuery, Postgres or a Rig-hosted warehouse you own. Each source lands in its own schema, and Rig’s engineers build and watch the pipelines and fix them when a source changes. Ingest is priced on the compute that runs your syncs, with no per-row or per-connector charge, and it is optional, so you can keep pipelines that already work.
The definitions sit above the pipeline. Rig Map holds table descriptions, joins and certified metrics in an open, OSI-compatible format, and Gateway controls which people and agents can query them. A merged education software group is replacing its Fivetran connections with Ingest in planned waves, checking each source for parity, including deleted records, before switching the old connection off.
Tools and approaches
- Fivetran
- A managed ingest service that extracts and loads data from many sources into a warehouse, priced on monthly active rows, with transformations measured separately in model runs.
- dbt
- A framework for transforming raw warehouse data with SQL select statements, version control and tests. It works alongside ingestion and BI tools, which makes it a common partner in ELT.
- Rig Ingest
- Managed scheduled syncs into your warehouse or a Rig-hosted one, set up and maintained by Rig’s engineers. Compare it with your current setup using the Fivetran cost calculator on the migration page.
Common questions
Is ELT always cheaper?
No. ELT moves cost into warehouse storage and compute and into the ingest tool’s pricing unit. It is often cheaper to maintain because the transforms are SQL that the data team can read, but compare the full bill for your own volumes.
Does Fivetran lock me into ELT?
Fivetran is built around extract and load into a destination, with transforms run there afterwards. If your business logic stays in the warehouse rather than in pipeline transformations, you can change ingest tools later without changing your metrics.
Where should business metrics be defined?
In the warehouse layer, in a semantic layer that BI tools, analysts and agents share. Keeping them out of the ingest layer means a pipeline change never changes a reported number.
Can I use ELT and ETL together?
Yes. Many teams mask regulated fields before load and use ELT for everything else. The rule that matters is that each metric has one definition, in one place.
What is reverse ETL?
Reverse ETL copies modelled data from the warehouse back into operational tools, such as a CRM. It reads from the same models, so it benefits from the same certified definitions.
Sources
Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.